On 18/06/2020 08:55, Chris Wilson wrote:
Quoting Tvrtko Ursulin (2020-06-18 08:40:25)
On 18/06/2020 08:14, Chris Wilson wrote:
Quoting Tvrtko Ursulin (2020-06-17 17:01:11)
From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Evaluation of userspace load balancing options was how this tool started
but since we have settled on doing it in the kernel.
Tomorrow we will want to update the tool for new engine interfaces and all
this legacy code will just be a distraction.
Rip out everything not related to explicit load balancing implemented via
context engine maps and adjust the workloads to use it.
Hmm, if this is on the table, should we also then restrict
load-balancing wsim to gen11+ so that we can use the timed loops rather
nop batches? That would be a huge selling point, and I'll just keep an
old checkout around for nop load balancing with all the trimmings.
That was my plan for the next step yes. Just taking your patch without
further changes would already make it work I think. But also at some
point I want to convert the engine selection (and engine naming in
descriptors) to class:instance.
Why do you need the nop/old balancing stuff? I would hope going forward
we only need to compare current balancing against any changes. So I'd
really like to remoev the userspace balancing stuff.
There are still some cases where i915 is beaten by plain old contexts,
usually that is a combination of semaphores and interrupt latency, but
some I just don't understand. There is still an uncomfortably large
variation between kernel releases, and comparing the regressions in
different balancers is useful to narrow down the problem.
You could create separate workloads to simulate "-b context" to a
degree? I really want to rip this out. Can you cut your losses and
forget it existed? :)
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
